Scholarlink

<p>Unsupervised few-shot learning (UFSL) is a special scenario of few-shot learning, where the training data are unlabeled. The main challenge of UFSL is to learn a good representation for obtaining similarity relations without the benefit of supervision. We propose a two-stage approach, named Masking Plus Refinement (MPR), to learn and enhance the similarity measurement in steps. In stage one, we obtain high-quality image representations through masked image reconstruction and numerically illustrate that the masking image operation helps overcome the overfitting problem of large models in few-shot learning. In stage two, we use the combination of metric loss and feature-invariant loss to refine the similarity measurement. Experimental results show that our MPR outperforms the state-of-the-art UFSL methods by 2%-4% in a 5-way 1-shot setting on the mini-ImageNet and tiered-ImageNet benchmarks while also achieving better results on higher-resolution images.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Masking plus refinement: A two-stage method for unsupervised few-shot learning

  • 期刊论文

  • 发表日期：

    2025年7月30日

查看全文
  • Jiajun Zhang,
  • Yangling Ma,
  • Yixin Luo

摘要

Unsupervised few-shot learning (UFSL) is a special scenario of few-shot learning, where the training data are unlabeled. The main challenge of UFSL is to learn a good representation for obtaining similarity relations without the benefit of supervision. We propose a two-stage approach, named Masking Plus Refinement (MPR), to learn and enhance the similarity measurement in steps. In stage one, we obtain high-quality image representations through masked image reconstruction and numerically illustrate that the masking image operation helps overcome the overfitting problem of large models in few-shot learning. In stage two, we use the combination of metric loss and feature-invariant loss to refine the similarity measurement. Experimental results show that our MPR outperforms the state-of-the-art UFSL methods by 2%-4% in a 5-way 1-shot setting on the mini-ImageNet and tiered-ImageNet benchmarks while also achieving better results on higher-resolution images.